Guia de início rápido: NVIDIA Jetson com Ultralytics YOLO26
Este guia explica como implementar o Ultralytics YOLO26 em dispositivos NVIDIA Jetson e inclui benchmarks de desempenho que mostram o que o YOLO26 pode fazer nesses dispositivos compactos e eficientes em termos de energia.
Atualizámos este guia com o mais recente NVIDIA Jetson AGX Thor Developer Kit, que oferece até 2070 FP4 TFLOPS de capacidade de computação de IA e 128 GB de memória, com potência configurável entre 40 W e 130 W. Ele oferece mais de 7,5 vezes a capacidade de computação de IA do NVIDIA Jetson AGX Orin, com eficiência energética 3,5 vezes superior para executar os modelos de IA mais populares.
Assista: Desbloqueie inferência mais rápida com o Ultralytics YOLO26 em dispositivos NVIDIA Jetson 🚀

Este guia foi testado com o NVIDIA Jetson AGX Thor Developer Kit (Jetson T5000) e o NVIDIA Jetson AGX Orin Developer Kit (64GB), com a versão estável mais recente do JetPack 7.2; com o NVIDIA Jetson Orin Nano Super Developer Kit, com a versão JP6.1 do JetPack; com o Seeed Studio reComputer J4012, baseado no NVIDIA Jetson Orin NX 16GB e com a versão JP6.0/JP5.1.3 do JetPack; e com o Seeed Studio reComputer J1020 v2, baseado no NVIDIA Jetson Nano 4GB e com a versão JP4.6.1 do JetPack. Espera-se que funcione com toda a gama de hardware NVIDIA Jetson, incluindo os dispositivos mais recentes e os mais antigos.
O que é o NVIDIA Jetson?#
O NVIDIA Jetson é uma série de placas de computação embarcadas projetadas para levar a computação acelerada de IA (inteligência artificial) a dispositivos de borda. Esses dispositivos compactos e potentes são construídos com base na arquitetura de GPU da NVIDIA e podem executar algoritmos complexos de IA e modelos de aprendizagem profunda diretamente no dispositivo, sem depender de recursos de computação em nuvem. As placas Jetson são frequentemente usadas em robótica, veículos autónomos, automação industrial e outras aplicações em que a inferência de IA precisa ser executada localmente, com baixa latência e alta eficiência. Além disso, essas placas são baseadas na arquitetura ARM64 e consomem menos energia do que os dispositivos tradicionais de computação com GPU.
Comparação das séries NVIDIA Jetson#
O NVIDIA Jetson AGX Thor é a versão mais recente da família NVIDIA Jetson, baseada na arquitetura NVIDIA Blackwell, que oferece um desempenho de AI significativamente superior ao das gerações anteriores. A tabela abaixo compara alguns dos dispositivos Jetson do ecossistema.
| Jetson AGX Thor (T5000) | Jetson AGX Orin 64GB | Jetson Orin NX 16GB | Jetson Orin Nano Super | Jetson AGX Xavier | Jetson Xavier NX | Jetson Nano | |
|---|---|---|---|---|---|---|---|
| Desempenho de IA | 2070 TFLOPS | 275 TOPS | 100 TOPS | 67 TOPS | 32 TOPS | 21 TOPS | 472 GFLOPS |
| GPU | GPU de arquitetura NVIDIA Blackwell com 2560 núcleos e 96 Tensor Cores | GPU de arquitetura NVIDIA Ampere com 2048 núcleos e 64 Tensor Cores | GPU de arquitetura NVIDIA Ampere com 1024 núcleos e 32 Tensor Cores | GPU de arquitetura NVIDIA Ampere com 1024 núcleos e 32 Tensor Cores | GPU de arquitetura NVIDIA Volta com 512 núcleos e 64 Tensor Cores | GPU de arquitetura NVIDIA Volta™ com 384 núcleos e 48 Tensor Cores | GPU de arquitetura NVIDIA Maxwell™ com 128 núcleos |
| Frequência máxima da GPU | 1.57 GHz | 1.3 GHz | 918 MHz | 1020 MHz | 1377 MHz | 1100 MHz | 921MHz |
| CPU | CPU Arm® Neoverse®-V3AE de 64 bits com 14 núcleos, 1 MB de L2 + 16 MB de L3 | CPU NVIDIA Arm® Cortex A78AE v8.2 de 64 bits com 12 núcleos, 3 MB de L2 + 6 MB de L3 | CPU NVIDIA Arm® Cortex A78AE v8.2 de 64 bits com 8 núcleos, 2 MB de L2 + 4 MB de L3 | CPU Arm® Cortex®-A78AE v8.2 de 64 bits e 6 núcleos, 1,5 MB de L2 + 4 MB de L3 | CPU NVIDIA Carmel Arm®v8.2 de 64 bits e 8 núcleos, 8 MB de L2 + 4 MB de L3 | CPU NVIDIA Carmel Arm®v8.2 de 64 bits e 6 núcleos, 6 MB de L2 + 4 MB de L3 | Processador Quad-Core Arm® Cortex®-A57 MPCore |
| Frequência máxima da CPU | 2.6 GHz | 2.2 GHz | 2.0 GHz | 1.7 GHz | 2.2 GHz | 1.9 GHz | 1.43GHz |
| Memória | 128 GB LPDDR5X de 256 bits, 273 GB/s | 64 GB LPDDR5 de 256 bits, 204,8 GB/s | 16 GB LPDDR5 de 128 bits, 102,4 GB/s | 8 GB LPDDR5 de 128 bits, 102 GB/s | 32 GB LPDDR4x de 256 bits, 136,5 GB/s | 8 GB LPDDR4x de 128 bits, 59,7 GB/s | 4 GB LPDDR4 de 64 bits, 25,6 GB/s |
O que é o NVIDIA JetPack?#
O NVIDIA JetPack SDK alimenta os módulos Jetson e oferece um ambiente de desenvolvimento completo para criar aplicações de IA aceleradas de ponta a ponta. O JetPack inclui o Jetson Linux com carregador de inicialização, kernel Linux, ambiente de desktop Ubuntu e um conjunto completo de bibliotecas para acelerar a computação com GPU, multimédia, gráficos e visão computacional. Também inclui exemplos, documentação e ferramentas para programadores, tanto para o computador anfitrião como para o kit de desenvolvimento, e é compatível com SDKs de nível superior, como o DeepStream para análise de vídeo em fluxo, o Isaac para robótica e o Riva para IA conversacional.
Instalar o JetPack na NVIDIA Jetson#
O primeiro passo depois de obteres um dispositivo NVIDIA Jetson é instalar o NVIDIA JetPack no dispositivo. Existem várias formas diferentes de instalar o JetPack em dispositivos NVIDIA Jetson.
- Para o JetPack 7.2 num kit de desenvolvimento oficial Jetson AGX Thor, AGX Orin ou Orin Nano, transfere a ISO unificada do Jetson, grava-a numa unidade USB e segue o guia de início rápido específico do dispositivo para AGX Thor, AGX Orin ou Orin Nano. A partir do JetPack 7.2, o Orin Nano já não utiliza uma imagem transferível para cartão SD; a instalação da ISO por USB instala o Jetson Linux no cartão microSD ou SSD NVMe do dispositivo.
- Se optares intencionalmente por usar o JetPack 6 num kit de desenvolvimento Jetson Orin Nano, segue as instruções da NVIDIA para atualizar o JetPack 6.x e usar o cartão SD.
- Se tiveres qualquer outro kit de desenvolvimento NVIDIA, podes instalar o JetPack no dispositivo com o SDK Manager.
- Se tiveres um dispositivo Seeed Studio reComputer J4012, podes instalar o JetPack no SSD incluído; se tiveres um dispositivo Seeed Studio reComputer J1020 v2, podes instalar o JetPack na eMMC/ SSD.
- Se tiveres qualquer outro dispositivo de terceiros equipado com um módulo NVIDIA Jetson, recomenda-se que sigas as instruções de instalação pela linha de comandos.
Para os métodos 1, 4 e 5 acima, depois de gravar a imagem do sistema e iniciar o dispositivo, executa sudo apt update && sudo apt install nvidia-jetpack -y no terminal do dispositivo para instalar todos os componentes restantes do JetPack necessários.
Suporte do JetPack por dispositivo Jetson#
A tabela abaixo destaca as versões do NVIDIA JetPack suportadas pelos diferentes dispositivos NVIDIA Jetson.
| JetPack 4 | JetPack 5 | JetPack 6 | JetPack 7 | |
|---|---|---|---|---|
| Jetson Nano | ✅ | ❌ | ❌ | ❌ |
| Jetson TX2 | ✅ | ❌ | ❌ | ❌ |
| Jetson Xavier NX | ✅ | ✅ | ❌ | ❌ |
| Jetson AGX Xavier | ✅ | ✅ | ❌ | ❌ |
| Jetson AGX Orin | ❌ | ✅ | ✅ | ✅ |
| Jetson Orin NX | ❌ | ✅ | ✅ | ✅ |
| Jetson Orin Nano | ❌ | ✅ | ✅ | ✅ |
| Jetson AGX Thor | ❌ | ❌ | ❌ | ✅ |
Início rápido com Docker#
A forma mais rápida de começar a usar o Ultralytics YOLO26 no NVIDIA Jetson é executar as imagens Docker pré-compiladas para Jetson. Consulta a tabela acima e escolhe a versão do JetPack correspondente ao teu dispositivo Jetson.
t=ultralytics/ultralytics:latest-jetson-jetpack4
sudo docker pull $t && sudo docker run -it --ipc=host --runtime=nvidia $tA imagem latest-nvidia-arm64 usa NVIDIA PyTorch 26.10, incluindo CUDA 13.4 e TensorRT 11.4. A NVIDIA indica o JetPack 7.1 para PyTorch na sua tabela de compatibilidade, mas o TensorRT 11 não é compatível com o JetPack, incluindo Thor. Usa esta imagem apenas para cargas de trabalho PyTorch num host compatível. Para exportações TensorRT no Jetson, usa a instalação nativa abaixo com o runtime TensorRT 10.x compatível com a tua versão do JetPack. O JetPack 7.2 no Thor ou no Orin requer validação separada do contentor. Reconstrói os engines para a GPU de destino e a versão do TensorRT.
Para imagens do JetPack 4, 5 e 6, continua para Utilizar o TensorRT na NVIDIA Jetson. A imagem do JetPack 7.1 acima destina-se apenas a cargas de trabalho PyTorch.
Começar com a instalação nativa#
Para uma instalação nativa sem Docker, consulta os passos abaixo.
Executar no JetPack 7.2#
Instalar o pacote Ultralytics#
Instala o pacote Ultralytics no Jetson. As dependências de exportação são instaladas automaticamente na primeira vez que exportas um modelo, por isso só precisas do pacote base neste caso. Este guia concentra-se nas exportações NVIDIA TensorRT, pois o TensorRT oferece o melhor desempenho nos dispositivos Jetson.
-
Atualizar a lista de pacotes, instalar o pip e atualizá-lo para a versão mais recente
sudo apt update sudo apt install python3-pip -y pip install -U pip -
Instalar o pacote pip
ultralyticspip install ultralytics -
Reiniciar o dispositivo
sudo reboot
Instalar o PyTorch e o Torchvision#
A instalação do ultralytics acima instala o Torch e o Torchvision. No entanto, esses dois pacotes instalados via pip não são compatíveis com dispositivos JetPack 7.2 que executam CUDA 13, por isso precisam ser instalados manualmente.
Instalar torch e torchvision de acordo com o JP7.2
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu130Instalar onnxruntime-gpu#
Utiliza uma wheel de ONNX Runtime GPU compatível com as versões do JetPack, Python e CUDA. As versões atuais no PyPI incluem wheels ARM64, mas não substituem os pacotes específicos do dispositivo em todas as versões do JetPack.
Descarrega e instala onnxruntime-gpu 1.24.0 com suporte a Python3.12.
pip install https://github.com/ultralytics/assets/releases/download/v0.0.0/onnxruntime_gpu-1.24.0-cp312-cp312-linux_aarch64.whlExecutar no JetPack 6.1#
Instalar o pacote Ultralytics#
Instala o pacote Ultralytics no Jetson. As dependências de exportação são instaladas automaticamente na primeira vez que exportas um modelo, por isso só precisas do pacote base neste caso. Este guia concentra-se nas exportações NVIDIA TensorRT, pois o TensorRT oferece o melhor desempenho nos dispositivos Jetson.
-
Atualizar a lista de pacotes, instalar o pip e atualizá-lo para a versão mais recente
sudo apt update sudo apt install python3-pip -y pip install -U pip -
Instalar o pacote pip
ultralyticspip install ultralytics -
Reiniciar o dispositivo
sudo reboot
Instalar o PyTorch e o Torchvision#
A instalação do ultralytics acima instala o Torch e o Torchvision. No entanto, esses dois pacotes instalados via pip não são compatíveis com a plataforma Jetson, que se baseia na arquitetura ARM64; por isso, tens de instalar manualmente um pacote wheel precompilado do PyTorch para pip e compilar ou instalar o Torchvision a partir do código-fonte.
Instalar torch 2.10.0 e torchvision 0.25.0 de acordo com o JP6.1
pip install https://github.com/ultralytics/assets/releases/download/v0.0.0/torch-2.10.0-cp310-cp310-linux_aarch64.whl
pip install https://github.com/ultralytics/assets/releases/download/v0.0.0/torchvision-0.25.0-cp310-cp310-linux_aarch64.whlVisita a página do PyTorch para Jetson para aceder a todas as versões do PyTorch disponíveis para as diferentes versões do JetPack. Para obter uma lista mais detalhada da compatibilidade entre PyTorch e Torchvision, visita a página de compatibilidade entre PyTorch e Torchvision.
Instala cuDSS para corrigir um problema de dependências com torch 2.10.0
wget https://developer.download.nvidia.com/compute/cudss/0.7.1/local_installers/cudss-local-tegra-repo-ubuntu2204-0.7.1_0.7.1-1_arm64.deb
sudo dpkg -i cudss-local-tegra-repo-ubuntu2204-0.7.1_0.7.1-1_arm64.deb
sudo cp /var/cudss-local-tegra-repo-ubuntu2204-0.7.1/cudss-*-keyring.gpg /usr/share/keyrings/
sudo apt-get update
sudo apt-get -y install cudssInstalar onnxruntime-gpu#
Utiliza uma wheel de ONNX Runtime GPU compatível com as versões do JetPack, Python e CUDA. As versões atuais no PyPI incluem wheels ARM64, mas não substituem os pacotes específicos do dispositivo em todas as versões do JetPack.
Podes encontrar todos os pacotes onnxruntime-gpu disponíveis — organizados por versão do JetPack, versão do Python e outros detalhes de compatibilidade — na matriz de compatibilidade do ONNX Runtime do Jetson Zoo.
Para o JetPack 6 com suporte para Python 3.10, podes instalar onnxruntime-gpu 1.23.0:
pip install https://github.com/ultralytics/assets/releases/download/v0.0.0/onnxruntime_gpu-1.23.0-cp310-cp310-linux_aarch64.whlEm alternativa, para onnxruntime-gpu 1.20.0:
pip install https://github.com/ultralytics/assets/releases/download/v0.0.0/onnxruntime_gpu-1.20.0-cp310-cp310-linux_aarch64.whlExecutar no JetPack 5.1.2#
Instalar o pacote Ultralytics#
Instala o pacote Ultralytics no Jetson. As dependências de exportação são instaladas automaticamente na primeira vez que exportas um modelo, por isso só precisas do pacote base neste caso. Este guia concentra-se nas exportações NVIDIA TensorRT, pois o TensorRT oferece o melhor desempenho nos dispositivos Jetson.
-
Atualizar a lista de pacotes, instalar o pip e atualizá-lo para a versão mais recente
sudo apt update sudo apt install python3-pip -y pip install -U pip -
Instalar o pacote pip
ultralyticspip install ultralytics -
Reiniciar o dispositivo
sudo reboot
Instalar o PyTorch e o Torchvision#
A instalação do ultralytics acima instala o Torch e o Torchvision. No entanto, esses dois pacotes instalados via pip não são compatíveis com a plataforma Jetson, que se baseia na arquitetura ARM64; por isso, tens de instalar manualmente um pacote wheel precompilado do PyTorch para pip e compilar ou instalar o Torchvision a partir do código-fonte.
-
Desinstalar o PyTorch e o Torchvision atualmente instalados
pip uninstall torch torchvision -
Instalar
torch 2.1.0etorchvision 0.16.2de acordo com o JP5.1.2pip install https://github.com/ultralytics/assets/releases/download/v0.0.0/torch-2.1.0a0+41361538.nv23.06-cp38-cp38-linux_aarch64.whl pip install https://github.com/ultralytics/assets/releases/download/v0.0.0/torchvision-0.16.2+c6f3977-cp38-cp38-linux_aarch64.whl
Visita a página do PyTorch para Jetson para aceder a todas as versões do PyTorch disponíveis para as diferentes versões do JetPack. Para obter uma lista mais detalhada da compatibilidade entre PyTorch e Torchvision, visita a página de compatibilidade entre PyTorch e Torchvision.
Instalar onnxruntime-gpu#
Utiliza uma wheel de ONNX Runtime GPU compatível com as versões do JetPack, Python e CUDA. As versões atuais no PyPI incluem wheels ARM64, mas não substituem os pacotes específicos do dispositivo em todas as versões do JetPack.
Podes encontrar todos os pacotes onnxruntime-gpu disponíveis — organizados por versão do JetPack, versão do Python e outros detalhes de compatibilidade — na matriz de compatibilidade do Jetson Zoo ONNX Runtime. Descarrega e instala onnxruntime-gpu 1.17.0 com suporte a Python3.8.
wget https://nvidia.box.com/shared/static/zostg6agm00fb6t5uisw51qi6kpcuwzd.whl -O onnxruntime_gpu-1.17.0-cp38-cp38-linux_aarch64.whl
pip install onnxruntime_gpu-1.17.0-cp38-cp38-linux_aarch64.whlonnxruntime-gpu atualiza automaticamente o NumPy para a versão mais recente; por isso, reinstala o NumPy 1.23.5 para corrigir o problema resultante:
pip install numpy==1.23.5
Utilizar o TensorRT na NVIDIA Jetson#
Entre todos os formatos de exportação de modelos compatíveis com Ultralytics, o TensorRT oferece o melhor desempenho de inferência nos dispositivos NVIDIA Jetson, o que faz dele a nossa principal recomendação para implementações no Jetson. Antes de exportar, instala os bindings Python do TensorRT fornecidos para a tua versão do JetPack e verifica-os com python3 -c "import tensorrt; print(tensorrt.__version__)". Mantém o runtime TensorRT 10.x compatível no JetPack 6/7; não o substituas pelo TensorRT 11. Para instruções de configuração e utilização avançada, consulta o nosso guia dedicado à integração do TensorRT.
Também podes exportar a partir do browser sem configurar localmente o ambiente de compilação. No separador Export do modelo na Ultralytics Platform, seleciona TensorRT e o dispositivo Jetson de destino. As seleções Thor são validadas em hardware Thor físico. As seis seleções Orin geram atualmente engines candidatos compilados para AGX-Orin; valida-os no modelo Orin pretendido antes da implementação.
O TensorRT cria e ajusta um mecanismo na GPU usada para a compilação. Faz corresponder a arquitetura da GPU e o ambiente de execução do TensorRT/CUDA aos do dispositivo de destino e valida cada mecanismo descarregado no dispositivo de implementação. Os modelos Orin com a mesma arquitetura não garantem automaticamente a portabilidade, e a calibração INT8 deve ser feita no dispositivo de destino para obter os melhores resultados.
Converter o modelo para TensorRT e executar a inferência#
O modelo YOLO26n no formato PyTorch é convertido para TensorRT para executar a inferência com o modelo exportado.
from ultralytics import YOLO
# Carregar um modelo YOLO26n PyTorch
model = YOLO("yolo26n.pt")
# Exportar o modelo para TensorRT
model.export(format="engine") # cria 'yolo26n.engine'
# Carregar o modelo TensorRT exportado
trt_model = YOLO("yolo26n.engine")
# Executar inferência
results = trt_model("https://ultralytics.com/images/bus.jpg")Utilizar o NVIDIA Deep Learning Accelerator (DLA)#
O NVIDIA Deep Learning Accelerator (DLA) é um componente de hardware especializado integrado nos dispositivos NVIDIA Jetson que otimiza a inferência de aprendizagem profunda para eficiência energética e desempenho. Ao transferir tarefas da GPU (libertando-a para processos mais exigentes), o DLA permite executar modelos com menor consumo de energia e, ao mesmo tempo, manter um elevado débito, algo ideal para sistemas integrados e aplicações de IA em tempo real.
A NVIDIA indica o TensorRT 10.7 como a última versão com suporte a DLA; o TensorRT 11.x não é compatível com DLA. Usa uma versão do TensorRT com suporte a DLA que seja compatível com a tua versão do JetPack. O TensorRT 11 não é compatível com o JetPack, incluindo as exportações que usam apenas a GPU.
Os seguintes dispositivos Jetson estão equipados com hardware DLA:
| Dispositivo Jetson | Núcleos DLA | Frequência máxima do DLA |
|---|---|---|
| Série Jetson AGX Orin | 2 | 1.6 GHz |
| Jetson Orin NX 16GB | 2 | 614 MHz |
| Jetson Orin NX 8GB | 1 | 614 MHz |
| Série Jetson AGX Xavier | 2 | 1.4 GHz |
| Série Jetson Xavier NX | 2 | 1.1 GHz |
from ultralytics import YOLO
# Carregar um modelo YOLO26n PyTorch
model = YOLO("yolo26n.pt")
# Exportar o modelo para TensorRT com DLA ativado (só funciona com FP16 ou INT8)
model.export(format="engine", device="dla:0", quantize=16) # dla:0 ou dla:1 corresponde aos núcleos DLA
# Carregar o modelo TensorRT exportado
trt_model = YOLO("yolo26n.engine")
# Executar inferência
results = trt_model("https://ultralytics.com/images/bus.jpg")Ao utilizar exportações DLA, algumas camadas podem não ser compatíveis com a execução no DLA e serão executadas na GPU. Esta alternativa pode introduzir latência adicional e afetar o desempenho geral da inferência. Por isso, o DLA não foi concebido principalmente para reduzir a latência de inferência em comparação com o TensorRT a funcionar inteiramente na GPU. O seu objetivo principal é, antes, aumentar o débito e melhorar a eficiência energética.
Benchmarks do NVIDIA Jetson com YOLO26#
A equipa do Ultralytics executou benchmarks do YOLO26 em 11 formatos de modelo diferentes, medindo velocidade e precisão: PyTorch, TorchScript, ONNX, OpenVINO, TensorRT, TF SavedModel, TF GraphDef, TF Lite, MNN, NCNN, ExecuTorch. Os benchmarks foram executados no NVIDIA Jetson AGX Thor Developer Kit, NVIDIA Jetson AGX Orin Developer Kit (64GB), NVIDIA Jetson Orin Nano Super Developer Kit e Seeed Studio reComputer J4012 com Jetson Orin NX 16GB, em precisão FP32 e com o tamanho de imagem de entrada padrão de 640.
Gráficos comparativos#
Embora todas as exportações de modelos funcionem no NVIDIA Jetson, o gráfico comparativo abaixo inclui apenas PyTorch, TorchScript e TensorRT. Você encontra os benchmarks de todas as exportações na seção após este gráfico.
Kit de desenvolvimento NVIDIA Jetson AGX Thor#
Kit de desenvolvimento NVIDIA Jetson AGX Orin (64GB)#
Kit de desenvolvimento NVIDIA Jetson Orin Nano Super#
NVIDIA Jetson Orin NX 16GB#
Tabelas de comparação detalhadas#
A tabela abaixo apresenta os resultados do benchmark de cinco modelos YOLO26 diferentes (YOLO26n, YOLO26s, YOLO26m, YOLO26l, YOLO26x) em 11 formatos diferentes (PyTorch, TorchScript, ONNX, OpenVINO, TensorRT, TF SavedModel, TF GraphDef, TF Lite, MNN, NCNN, ExecuTorch), indicando o estado, o tamanho, a métrica mAP50-95(B) e o tempo de inferência de cada combinação.
Kit de desenvolvimento NVIDIA Jetson AGX Thor#
| Formato | Status | Tamanho em disco (MB) | mAP50-95(B) | Tempo de inferência (ms/im) |
|---|---|---|---|---|
| PyTorch | ✅ | 5.3 | 0.4798 | 7.39 |
| TorchScript | ✅ | 9.8 | 0.4789 | 4.21 |
| ONNX | ✅ | 9.5 | 0.4767 | 6.58 |
| OpenVINO | ✅ | 10.1 | 0.4794 | 17.50 |
| TensorRT (FP32) | ✅ | 13.9 | 0.4791 | 1.90 |
| TensorRT (FP16) | ✅ | 7.6 | 0.4797 | 1.39 |
| TensorRT (INT8) | ✅ | 6.5 | 0.4273 | 1.52 |
| TF SavedModel | ✅ | 25.7 | 0.4764 | 47.24 |
| TF GraphDef | ✅ | 9.5 | 0.4764 | 45.98 |
| TF Lite | ✅ | 9.9 | 0.4764 | 182.04 |
| MNN | ✅ | 9.4 | 0.4784 | 21.83 |
Benchmark realizado com Ultralytics 8.4.7
O tempo de inferência não inclui o pré/pós-processamento.
Kit de desenvolvimento NVIDIA Jetson AGX Orin (64GB)#
| Formato | Status | Tamanho em disco (MB) | mAP50-95(B) | Tempo de inferência (ms/im) |
|---|---|---|---|---|
| PyTorch | ✅ | 5.3 | 0.4790 | 11.58 |
| TorchScript | ✅ | 9.8 | 0.4770 | 4.60 |
| ONNX | ✅ | 9.5 | 0.4770 | 9.87 |
| OpenVINO | ✅ | 9.6 | 0.4820 | 28.80 |
| TensorRT (FP32) | ✅ | 11.5 | 0.4770 | 4.18 |
| TensorRT (FP16) | ✅ | 7.9 | 0.4789 | 2.62 |
| TensorRT (INT8) | ✅ | 5.4 | 0.4640 | 2.30 |
| TF SavedModel | ✅ | 24.6 | 0.4760 | 71.10 |
| TF GraphDef | ✅ | 9.5 | 0.4760 | 70.02 |
| TF Lite | ✅ | 9.9 | 0.4760 | 227.94 |
| MNN | ✅ | 9.4 | 0.4760 | 32.46 |
| NCNN | ✅ | 9.3 | 0.4810 | 29.93 |
Benchmark realizado com Ultralytics 8.4.32
O tempo de inferência não inclui o pré/pós-processamento.
Kit de desenvolvimento NVIDIA Jetson Orin Nano Super#
| Formato | Status | Tamanho em disco (MB) | mAP50-95(B) | Tempo de inferência (ms/im) |
|---|---|---|---|---|
| PyTorch | ✅ | 5.3 | 0.4790 | 15.60 |
| TorchScript | ✅ | 9.8 | 0.4770 | 12.60 |
| ONNX | ✅ | 9.5 | 0.4760 | 15.76 |
| OpenVINO | ✅ | 9.6 | 0.4820 | 56.23 |
| TensorRT (FP32) | ✅ | 11.3 | 0.4770 | 7.53 |
| TensorRT (FP16) | ✅ | 8.1 | 0.4800 | 4.57 |
| TensorRT (INT8) | ✅ | 5.3 | 0.4490 | 3.80 |
| TF SavedModel | ✅ | 24.6 | 0.4760 | 118.33 |
| TF GraphDef | ✅ | 9.5 | 0.4760 | 116.30 |
| TF Lite | ✅ | 9.9 | 0.4760 | 286.00 |
| MNN | ✅ | 9.4 | 0.4760 | 68.77 |
| NCNN | ✅ | 9.3 | 0.4810 | 47.50 |
Benchmark realizado com Ultralytics 8.4.33
O tempo de inferência não inclui o pré/pós-processamento.
NVIDIA Jetson Orin NX 16GB#
| Formato | Status | Tamanho em disco (MB) | mAP50-95(B) | Tempo de inferência (ms/im) |
|---|---|---|---|---|
| PyTorch | ✅ | 5.3 | 0.4799 | 13.90 |
| TorchScript | ✅ | 9.8 | 0.4787 | 11.60 |
| ONNX | ✅ | 9.5 | 0.4763 | 14.18 |
| OpenVINO | ✅ | 9.6 | 0.4819 | 40.19 |
| TensorRT (FP32) | ✅ | 11.4 | 0.4770 | 7.01 |
| TensorRT (FP16) | ✅ | 8.0 | 0.4789 | 4.13 |
| TensorRT (INT8) | ✅ | 5.5 | 0.4489 | 3.49 |
| TF SavedModel | ✅ | 24.6 | 0.4764 | 92.34 |
| TF GraphDef | ✅ | 9.5 | 0.4764 | 92.06 |
| TF Lite | ✅ | 9.9 | 0.4764 | 254.43 |
| MNN | ✅ | 9.4 | 0.4760 | 48.55 |
| NCNN | ✅ | 9.3 | 0.4805 | 34.31 |
Benchmark realizado com Ultralytics 8.4.33
O tempo de inferência não inclui o pré/pós-processamento.
Explore mais iniciativas de benchmarking da Seeed Studio executadas em diferentes versões do hardware NVIDIA Jetson.
Reproduza nossos resultados#
Para reproduzir os benchmarks da Ultralytics acima em todos os formatos de exportação, execute este código:
from ultralytics import YOLO
# Carregar um modelo YOLO26n PyTorch
model = YOLO("yolo26n.pt")
# Benchmarke a velocidade e a precisão do YOLO26n no conjunto de dados COCO128 para todos os formatos de exportação
results = model.benchmark(data="coco128.yaml", imgsz=640)Observe que os resultados do benchmark podem variar conforme a configuração exata de hardware e software de um sistema, bem como a carga de trabalho atual do sistema no momento da execução dos benchmarks. Para obter resultados mais confiáveis, use um conjunto de dados com um grande número de imagens, por exemplo, data='coco.yaml' (5000 imagens de validação).
Práticas recomendadas ao usar NVIDIA Jetson#
Segue estas práticas recomendadas para obter o máximo desempenho do YOLO26 no NVIDIA Jetson.
-
Ative o modo de potência MAX
Ativar o modo de potência MAX no Jetson garante que todos os núcleos da CPU e da GPU estejam ativos.
sudo nvpmodel -m 0 -
Ative os Jetson Clocks
Ativar os Jetson Clocks garante que todos os núcleos da CPU e da GPU funcionem na frequência máxima.
sudo jetson_clocks -
Instale o aplicativo Jetson Stats
Usa a aplicação jetson-stats para monitorizar as temperaturas dos componentes do sistema, ver a utilização da CPU, GPU e RAM, alterar os modos de potência, definir as frequências máximas e consultar as informações do JetPack.
sudo apt update sudo pip install jetson-stats sudo reboot jtop
Dicas de otimização de memória para NVIDIA Jetson#
A memória disponível costuma ser o fator limitante nos dispositivos Jetson, principalmente nas variantes com menos memória, como o Jetson Orin Nano (8 GB) ou o Orin NX 8 GB. As dicas abaixo são mudanças práticas e de baixo risco que, em conjunto, podem liberar várias centenas de megabytes e permitir executar modelos YOLO maiores ou oferecer suporte a mais cargas de trabalho paralelas. Para uma visão abrangente, consulte o artigo da NVIDIA sobre como maximizar a eficiência da memória no Jetson.
1. Mude para a inicialização sem interface gráfica (headless)#
Se o Jetson estiver conectado por SSH ou funcionando como um dispositivo de produção sem monitor conectado, eliminar o ambiente de desktop e o servidor de exibição pode liberar até 865 MB de RAM:
sudo systemctl set-default multi-user.target
sudo rebootPara restaurar o ambiente de desktop mais tarde:
sudo systemctl set-default graphical.target
sudo reboot2. Desative serviços do sistema não utilizados#
Serviços em segundo plano não essenciais (Bluetooth, gerenciadores de conectividade, daemons de hardware não utilizados) consomem cerca de 32 MB no total. Liste os serviços ativos e desative tudo de que sua implantação não precisa:
# List running services
systemctl list-units --type=service --state=running
# Disable a service
sudo systemctl disable SERVICE_NAME3. Analise o uso de memória#
Antes de otimizar, identifique quais processos estão realmente consumindo RAM. procrank ordena os processos por PSS (Proportional Set Size), que reflete o uso de memória real de cada processo com mais precisão do que RSS (Resident Set Size, o total de páginas de RAM física mapeadas por um processo, incluindo páginas compartilhadas com outros processos):
git clone https://github.com/csimmonds/procrank_linux.git
cd procrank_linux && make
sudo ./procrankPara ver as alocações de GPU e NvMap (pipeline CUDA/vídeo) por processo:
sudo cat /sys/kernel/debug/nvmap/iovmm/clients4. Execute inferência sem monitor em produção#
Em pipelines de inferência que não precisam de pré-visualização ao vivo, desativar componentes relacionados à exibição (Tiler, OSD, DisplaySink) pode economizar mais de 200 MB apenas no pipeline. Com Ultralytics YOLO, desative o visualizador e grave os resultados em disco:
from ultralytics import YOLO
model = YOLO("yolo26n.engine")
# show=False impede qualquer janela de exibição; save=True grava a saída anotada em disco
results = model.predict(source="video.mp4", show=False, save=True)Impacto acumulado#
| Otimização | Memória aproximada liberada |
|---|---|
| Desativar a interface gráfica do desktop | ~865 MB |
| Desativar serviços do sistema operacional não utilizados | ~32 MB |
| Pipeline de inferência sem interface gráfica (sem exibição) | ~200+ MB |
| Total (ganhos fáceis) | ~1 GB+ |
Combinar essas mudanças é especialmente valioso ao usar modelos TensorRT INT8 em dispositivos com memória limitada — pode ser a diferença entre conseguir ou não manter na memória uma variante maior do modelo.
Próximas etapas#
Para saber mais e obter suporte, consulte a documentação do Ultralytics YOLO26.
Perguntas frequentes#
Para implementar o Ultralytics YOLO26 em dispositivos NVIDIA Jetson, primeiro grava a imagem do NVIDIA JetPack SDK no teu dispositivo Jetson. Em seguida, usa uma imagem Docker pré-compilada para uma configuração rápida ou instala manualmente os pacotes necessários.
Os modelos YOLO26 foram avaliados em dispositivos NVIDIA Jetson AGX Thor, AGX Orin, Orin Nano Super e Orin NX, e o formato TensorRT oferece o melhor desempenho de inferência. A secção Tabelas de comparação detalhadas apresenta mAP50-95 e o tempo de inferência de cada modelo e formato de exportação.
O TensorRT é recomendado para implantar modelos YOLO26 no NVIDIA Jetson porque oferece a inferência mais rápida entre todos os formatos de exportação, otimizando o modelo para a GPU do Jetson.
Para instalar PyTorch e Torchvision em NVIDIA Jetson, primeiro desinstale as versões existentes que possam ter sido instaladas via pip. Em seguida, instale manualmente as versões compatíveis de PyTorch e Torchvision para a arquitetura ARM64 do Jetson. As instruções detalhadas desse processo estão na seção Instale PyTorch e Torchvision.
Para maximizar o desempenho de YOLO26 em NVIDIA Jetson, siga estas práticas recomendadas:
- Ative o modo de potência MAX para utilizar todos os núcleos de CPU e GPU.
- Ative os Jetson Clocks para operar todos os núcleos na frequência máxima.
- Instale o aplicativo Jetson Stats para monitorar as métricas do sistema.
Para ver os comandos e mais detalhes, consulte a seção Práticas recomendadas ao usar NVIDIA Jetson.
A RAM disponível costuma ser o gargalo nos dispositivos Jetson com menos memória. Três ganhos fáceis que, juntos, podem liberar mais de 1 GB:
- Mude para a inicialização sem interface gráfica (
sudo systemctl set-default multi-user.target) para eliminar a interface gráfica do desktop (~865 MB liberados). - Desative serviços não utilizados, como Bluetooth ou gerenciadores de conectividade (~32 MB liberados).
- Execute inferência sem monitor definindo
show=Falsena chamada YOLOpredict, o que evita alocar memória para o pipeline de exibição (~200+ MB liberados).
Use
procrankpara analisar o uso de RAM por processo esudo cat /sys/kernel/debug/nvmap/iovmm/clientspara inspecionar as alocações de GPU. Consulte a seção Dicas de otimização de memória para ver todos os detalhes.- Mude para a inicialização sem interface gráfica (
O TensorRT 10.3.0 incluído no JetPack 6 tem um problema conhecido que impede a criação de mecanismos INT8 com saídas sem NMS (
nms=False) ou com NMS incorporado (nms=True). Quando o Ultralytics detecta essa combinação, seleciona a cabeça de um para muitos para exportações sem NMS e interrompe as exportaçõesnms=Truecom um erro.Para restaurar essas exportações INT8, atualiza o TensorRT para uma versão mais recente (por exemplo, 10.7.0+):
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/arm64/cuda-keyring_1.1-1_all.deb sudo dpkg -i cuda-keyring_1.1-1_all.deb sudo apt-get update sudo apt-get install -y tensorrtDepois de atualizar, volta a executar a exportação. Para mais detalhes, consulta o problema #23841 no GitHub e o problema #26655.
Comentários194
Faz uma pergunta, partilha os teus resultados ou ajuda alguém.
LG973641114
您好,如何用c++在jetson上部署YOLO26,有相关例子吗
1pknowler
Struggling with this on a new Jetson Orin Nano JP6.1. All good following the steps and installs to give torch.cuda.is_available as TRUE. The TensorRT export fails though with a number of dependency errors and I can’t re-install it with anything. Any pointers please?
1
Glenn JocherUltralytics@pknowler it seems like the dependency errors during TensorRT export might stem from compatibility issues. Please ensure you’ve installed the required JetPack-compatible versions of PyTorch, Torchvision, and TensorRT as outlined in the Jetson Quick Start Guide. Additionally, verify that
onnxruntime-gpuandcuSPARSELtare correctly installed per the guide. If the issue persists, you can try reinstalling TensorRT using the NVIDIA SDK Manager for your JetPack version. Let me know if you need further clarification!pknowler
Thank you. I will try again working through all of the steps and let you know how it goes!
pknowler
I’ve tried again but this as far as I get… I have followed everything carefully.
WARNING ⚠️ TensorRT requires GPU export, automatically assigning device=0 Ultralytics 8.3.57 🚀 Python-3.10.16 torch-2.5.0a0+872d972e41.nv24.08 CUDA:0 (Orin, 7620MiB) YOLO11n summary (fused): 238 layers, 2,616,248 parameters, 0 gradients, 6.5 GFLOPs
PyTorch: starting from 'yolo11n.pt' with input shape (1, 3, 640, 640) BCHW and output shape(s) (1, 84, 8400) (5.4 MB)
ONNX: starting export with onnx 1.17.0 opset 19... ONNX: slimming with onnxslim 0.1.46... ONNX: export success ✅ 3.1s, saved as 'yolo11n.onnx' (10.2 MB) requirements: Ultralytics requirement ['tensorrt>7.0.0,!=10.1.0'] not found, attempting AutoUpdate... error: subprocess-exited-with-error
× python setup.py egg_info did not run successfully. │ exit code: 1 ╰─> [6 lines of output] Traceback (most recent call last): File "", line 2, in File "", line 34, in File "/tmp/pip-install-d0ljzl60/tensorrt-cu12_db0552ac431148489703f6f75df6764f/setup.py", line 71, in raise RuntimeError("TensorRT does not currently build wheels for Tegra systems") RuntimeError: TensorRT does not currently build wheels for Tegra systems [end of output]
note: This error originates from a subprocess, and is likely not a problem with pip. error: metadata-generation-failed
× Encountered error while generating package metadata. ╰─> See above for output.
note: This is an issue with the package mentioned above, not pip. hint: See above for details. Retry 1/2 failed: Command 'pip install --no-cache-dir "tensorrt>7.0.0,!=10.1.0" ' returned non-zero exit status 1. error: subprocess-exited-with-error
× python setup.py egg_info did not run successfully. │ exit code: 1 ╰─> [6 lines of output] Traceback (most recent call last): File "", line 2, in File "", line 34, in File "/tmp/pip-install-782zjqf9/tensorrt-cu12_44d261aea7c5489db3e9ce15cec8ee64/setup.py", line 71, in raise RuntimeError("TensorRT does not currently build wheels for Tegra systems") RuntimeError: TensorRT does not currently build wheels for Tegra systems [end of output]
note: This error originates from a subprocess, and is likely not a problem with pip. error: metadata-generation-failed
× Encountered error while generating package metadata. ╰─> See above for output.
note: This is an issue with the package mentioned above, not pip. hint: See above for details.
pknowler
And I have TensorRT 10.7.0.23-1+cuda12.6
Glenn JocherUltralytics@pknowler it seems your TensorRT version is not compatible with the Jetson's ARM64 architecture. TensorRT does not provide pre-built Python wheels for Tegra systems. Please ensure you have TensorRT installed via the NVIDIA SDK Manager as part of the JetPack installation, and confirm compatibility with your JetPack version. The Jetson Quick Start Guide might also help clarify setup specifics. Let me know if further assistance is needed!
pknowler
@glenn-jocher thank you. I have found the issue now. I am using miniconda and for some reason TensorRT Python bindings are not available in conda envs. So I can import Torch with no problem but the import TensorRT fails. I have re-installed without conda and all works as expected! 👍🏻
RasoulZardeyni
Which one reduce inference time using TensorRT or deepstream?
1
Paula DerrengerUltralytics(editado)In terms of reducing inference time on your NVIDIA Jetson device, using TensorRT is generally recommended. TensorRT is specifically designed to provide high performance for NVIDIA hardware by optimizing deep learning models for inference. It accelerates models by fusing layers, pruning unused layers, and quantizing variables, among other optimizations. 🚀
DeepStream, on the other hand, is more focused on streaming analytics and can utilize TensorRT-optimized models. For most cases, directly using TensorRT for model optimization is your go-to for minimizing inference times.
Here's a quick example of exporting a YOLO model to TensorRT and running inference, tailored for your convenience:
from ultralytics import YOLO # Load your YOLO model model = YOLO('yolo26n.pt') # Export to TensorRT format model.export(format='engine') # Load the TensorRT model trt_model = YOLO('yolo26n.engine') # Run inference using the TensorRT model results = trt_model('your_image_or_video.jpg')Feel free to explore more about deploying on NVIDIA Jetson with our Quick Start Guide for NVIDIA Jetson. Happy inferencing! 🌟
2RasoulZardeyni
Thanks!
1
Paula DerrengerUltralytics(editado)Hey there! 😊 It looks like you're interested in speeding up your inference times. Between TensorRT and DeepStream on NVIDIA Jetson devices, TensorRT typically offers more direct optimization for reducing inference time for individual models by providing lower-level optimizations and quantization options.
DeepStream is a comprehensive framework for building end-to-end video analytics applications, which utilizes TensorRT under the hood for model inference. DeepStream excels in scenarios involving video streams, multi-stream input, and complex pipelines involving multiple AI models and analytics. If your primary goal is solely to minimize inference time for a specific model, direct optimization via TensorRT can provide more focused improvements.
A quick example of exporting a YOLO model to TensorRT:
from ultralytics import YOLO # Load your YOLO model model = YOLO('yolo26n.pt') # Export to TensorRT format model.export(format='engine') # This creates 'yolo26n.engine' # Done! Now you can use 'yolo26n.engine' for faster inference.However, if your application benefits from a broader video analytics pipeline, incorporating DeepStream into your workflow can optimize and streamline these processes on top of efficiently handling inference with TensorRT optimized models. 🚀
Hope this gives you a clearer picture! If you have more questions, feel free to ask.
1
Yufeng Wu(editado)Hi, ultralytics team. I'm trying yolo26n-pose on Jetson orin agx 64gb and have this problem: I tried building engine with dynamic=true but failed, with the log saying ForeignNode[/model.23/Sigmoid.../model.23/Flatten] (Myelin) Skipping tactic ... Incompatible effective shapes in operation /model_23/Sub between: Unsqueeze_18_output_0 f32[1,2,33600] and Slice_output_0 f32[1,2,21504] Could not find any implementation for node {ForeignNode[/model.23/Sigmoid.../model.23/Flatten]} Is there a way to fix it?
Glenn JocherUltralyticsThis indicates a TensorRT dynamic-shape profile failure in the pose head. As a workaround, export a fixed-resolution engine, e.g.
yolo export model=yolo26n-pose.pt format=engine imgsz=640 batch=1 dynamic=False; build separate engines if multiple resolutions are required. If dynamic shapes are essential, please open a new issue with the full text log,yolo checksoutput, and JetPack/TensorRT/Ultralytics versions so we can reproduce it; see the TensorRT export guide.
Haotian Gong(editado)Hi, everyone,
I'd like to flag two issues with the Jetson installation guide that may need updating.
1. PyTorch 2.10.0 / Torchvision 0.25.0 doesn't work on JetPack 6.1
Following the "Start with naive installation" and "Run on JetPack 6.1" sections, I installed Pytorch 2.10.0 and Torchvision 0.25.0 as instructed.
Running my script then throws:
ImportError: libcublas.so.13: cannot open shared object file: No such file or directorySo I have to install Pytorch 2.8.0 with the guide:
https://forums.developer.nvidia.com/t/jetson-orin-nano-issues-running-yolo-ultralytics-due-to-pytorch-torchvision-compatibility-jetpack-6-x-cuda-12-6/365683/9And this version of Pytorch works. So maybe is the guide on Jetson obsolete or mistaken?
2. CSI camera requires the JetPack-bundled OpenCV, not the PyPI one
On Jetson, using a CSI camera (via
nvarguscamerasrc) requires the OpenCV that ships with JetPack, because the PyPIopencv-pythonwheel on aarch64 is built without GStreamer support.However,
pip install ultralyticspulls inopencv-pythonas a dependency, which shadows the JetPack-bundled version and silently breaks CSI camera pipelines.pip uninstall -y opencv-python opencv-python-headless sudo apt install -y python3-opencv # JetPack-bundled OpenCV with GStreamerThanks!
Lakshantha DissanayakeUltralyticsHi @ZephyrKeXiner Could you please double check whether you are using JetPack 6.1 system with CUDA 12.6? I have just verified on my JetPack 6.1 system with CUDA 12.6 with
torch-2.10.0andtorchvision-0.25.0as mentioned in the Ultralytics Jetson Guide and it works well.For JetPack 6.1, output of
cat /etc/nv_tegra_releaseshould look like below:(env) nvidia@ubuntu:~/yolo26/ultralytics$ cat /etc/nv_tegra_release # R36 (release), REVISION: 4.3, GCID: 38968081, BOARD: generic, EABI: aarch64, DATE: Wed Jan 8 01:49:37 UTC 2025 # KERNEL_VARIANT: oot TARGET_USERSPACE_LIB_DIR=nvidia TARGET_USERSPACE_LIB_DIR_PATH=usr/lib/aarch64-linux-gnu/nvidiaCould you also share yours please? Thanks!
cc: @raimbekovm
1
Haotian GongHi @lakshanthad,
My output of
cat /etc/nv_tegra_releaseis below:$ cat /etc/nv_tegra_release # R36 (release), REVISION: 5.0, GCID: 43688277, BOARD: generic, EABI: aarch64, DATE: Fri Jan 16 03:50:45 UTC 2026 # KERNEL_VARIANT: oot TARGET_USERSPACE_LIB_DIR=nvidia TARGET_USERSPACE_LIB_DIR_PATH=usr/lib/aarch64-linux-gnu/nvidiaI notice that the revision is 5.0 not 4.3. Maybe does it cause the error?
smbunn(editado)
I would have thought the Jetson Orin Nano could do more than 5 streams, the deepstream examples with cars handles 16 streams of 1080p at what appears to be full framerate.
As a control systems engineer by training, edge devices seem a much better way to go so I prefer your jetson design. Even if one Jetson has an issue you are still processing 45 streams whereas a problem on your centralised server in option 2 takes everything down.
Are you going to set up an MQTT feed for alerts from each stream? Maybe consider secondary processing and a buffer, so the edge device detects what it thinks is an issue and a centralised server grabs 5 seconds of video of the incident and double checks.
Simon BUNN
General
From: Yash Patel @.> Sent: Friday, 22 May 2026 4:00 am To: ultralytics/ultralytics @.> Cc: Bunn, Simon @.>; Mention @.> Subject: Re: [ultralytics/ultralytics] guides/nvidia-jetson/ (Discussion #9942)
Hi, I am designing an AI-based CCTV surveillance system for a residential society with 50 RTSP camera streams. My current architecture uses multiple NVIDIA Jetson Orin Nano Super Developer Kit devices for distributed inference processing. I ZjQcmQRYFpfptBannerStart Attention The sender's identity could not be verified and someone may be impersonating the sender. L'identité de l'expéditeur n'a pas pu être vérifiée et quelqu'un pourrait se faire passer pour l'expéditeur.
ZjQcmQRYFpfptBannerEnd
Hi, I am designing an AI-based CCTV surveillance system for a residential society with 50 RTSP camera streams. My current architecture uses multiple NVIDIA Jetson Orin Nano Super Developer Kit devices for distributed inference processing. I estimate that one Jetson Orin Nano Super (8GB) can process around 5 RTSP streams at usable real-time FPS. This requires approximately 10 Jetson devices for the full deployment. Each device costs around ₹33,000 (~USD 395). The total estimated hardware cost is ₹330,000 (~USD 3,950). I want to compare this setup against a centralized Linux-based custom PC/server with NVIDIA GPU acceleration under the same budget. The comparison focuses on AI inference performance, scalability, reliability, and 24/7 operational efficiency. The workload mainly includes multi-stream YOLO/object detection inference on live RTSP feeds. I want to determine which architecture provides the best computation-to-cost ratio for this surveillance use case.
— Reply to this email directly, view it on GitHubhttps://urldefense.com/v3/__https://github.com/orgs/ultralytics/discussions/9942*discussioncomment-17006943__;Iw!!ElGdukoduuk!URiPjKkgDoMT_k2Y6pEd3XMEBGMDjhwgtQmwWXR9F4R4L7ymNloZD79NanOjYVg9YYmaARqyZGbgFv7U15KzgdcU$, or unsubscribehttps://urldefense.com/v3/__https://github.com/notifications/unsubscribe-auth/ABAAMILYUP7PPODNHDR3HC3434RYLAVCNFSM6AAAAABGAK2IGGVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTOMBQGY4TIMY__;!!ElGdukoduuk!URiPjKkgDoMT_k2Y6pEd3XMEBGMDjhwgtQmwWXR9F4R4L7ymNloZD79NanOjYVg9YYmaARqyZGbgFv7U1zjhi4YU$. Triage notifications on the go with GitHub Mobile for iOShttps://urldefense.com/v3/__https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675__;!!ElGdukoduuk!URiPjKkgDoMT_k2Y6pEd3XMEBGMDjhwgtQmwWXR9F4R4L7ymNloZD79NanOjYVg9YYmaARqyZGbgFv7U13gOq1qb$ or Androidhttps://urldefense.com/v3/__https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign*3Dnotification-email*26utm_medium*3Demail*26utm_source*3Dgithub__;JSUlJSU!!ElGdukoduuk!URiPjKkgDoMT_k2Y6pEd3XMEBGMDjhwgtQmwWXR9F4R4L7ymNloZD79NanOjYVg9YYmaARqyZGbgFv7U1zOl3XaF$. You are receiving this because you were mentioned.Message ID: @.***>
Glenn JocherUltralytics5streams on an Orin Nano Super is just a conservative YOLO/Python estimate; with a headless TensorRT or DeepStream pipeline, MAX power mode, andjetson_clocks, higher 1080p stream counts are definitely possible and depend heavily on the model size, input resolution, decode/encode, and any tracking or alert logic you add, so please drop your exactmodel,imgsz, FPS target, and RTSP resolution here and we can help size it more accurately using the Jetson deployment guide and DeepStream Jetson guide.
yashpatel2010
Hi, I am designing an AI-based CCTV surveillance system for a residential society with 50 RTSP camera streams. My current architecture uses multiple NVIDIA Jetson Orin Nano Super Developer Kit devices for distributed inference processing. I estimate that one Jetson Orin Nano Super (8GB) can process around 5 RTSP streams at usable real-time FPS. This requires approximately 10 Jetson devices for the full deployment. Each device costs around ₹33,000 (~USD 395). The total estimated hardware cost is ₹330,000 (~USD 3,950). I want to compare this setup against a centralized Linux-based custom PC/server with NVIDIA GPU acceleration under the same budget. The comparison focuses on AI inference performance, scalability, reliability, and 24/7 operational efficiency. The workload mainly includes multi-stream YOLO/object detection inference on live RTSP feeds. I want to determine which architecture provides the best computation-to-cost ratio for this surveillance use case.
RotatingGiraffe
Hi there, I am trying to run a fine-tuned YOLO26 based model for an OBB task on a Xavier NX running JetPack 5. I am using your container (ultralytics/ultralytics:latest-jetson-jetpack5) and trying to compile from PyTorch to TensorRT. I get the error: "mymodel.pt requires numpy>=1.26.1, however numpy==1.23.5 is installed."
Previously I have successfully the file on my host machine from PT to ONNX, copied ONNX to Jetson and then used tensorrtexec to compile to Engine. But the resulting models did not create any bounding boxes.
I wanted to ask if JetPack 5 even supports Yolo26 or if the model contains some incompatible layers? YOLO11 OBB works fine which I find extra-weird. Or is there likely an error somewhere else in my script?
Glenn JocherUltralyticsThanks for posting this here—JetPack 5 on Xavier NX is supported for YOLO26/TensorRT in our Jetson guide, so the
numpy>=1.26.1error points more to an environment/checkpoint mismatch than unsupported OBB layers; please paste the full text traceback andyolo checksoutput, and note that an engine built withtensorrtexeccan still produce empty OBB results if the Ultralytics export metadata/postprocessing is not preserved.RotatingGiraffe
Hi there, I tried to create a minimal reproducible example for you, based on what I am doing. I hope it helps with debugging.
The yolo checks you asked for is in the script output, along with the export script. I've also included the bit where I export via trtexex just for the sake of completion - it might very well be possible that I omit required metadata with this. My prefered goal would be that everything works via the Ultralytics Python API, but I had to search for workaround due to the NumPy error when compiling PT files.
I am now using the Ultralytics base image in the MRE which I hoped would be easier to debug for us. The NumPy error persists on the new image. Previously I was using something built upon l4t-ml.
Additionally, ONNX fails with an error ( `GLIBCXX_3.4.29' not found I think it was) but that is not a concern to me since I engine would be my prefered way to go. I have not installed any additional packages and kept it very vanilla (see dockerfile). The MRE also weirdly produced a new error during engine inference, but I'd say lets just focus on getting the Ultralytics native pipeline to work and then maybe all the other issues were just from my workarounds and disappear.
Execution Command
sudo docker compose up --buildDirectory Structure
obb-test-repro/ ┣ 📂data/ ┃ ┗ 📜mydataset.yaml ┣ 📂models/ ┃ ┣ 📜my-yolo26m-model.onnx ┃ ┗ 📜my-yolo26m-model.pt ┣ 📂src/ ┃ ┗ 📜main.py ┣ 📜docker-compose.yml ┣ 📜Dockerfile ┗ 📜requirements.txtDockerfile
FROM ultralytics/ultralytics:latest-jetson-jetpack5 WORKDIR /app # Copy the test app COPY src/ ./src/ # Run the test CMD ["python3", "src/main.py"]docker-compose.yml
version: '3.8' services: obb-test-jp5: image: obb-test:latest container_name: obb_test_jp5 build: context: . dockerfile: Dockerfile runtime: nvidia network_mode: host ipc: host restart: "no" volumes: - ./models:/app/models - ./data:/app/data environment: - NVIDIA_VISIBLE_DEVICES=all - NVIDIA_DRIVER_CAPABILITIES=allmain.py
""" Minimal OBB model test script. Behavior: - Runs `yolo checks` to output environment details. - Uses files in `/app/models` (container path mapped from `obb-test/models`). - Expects two files (hard-coded): - `/app/models/my-yolo26m-model.pt` (optional) — used for Ultralytics native export - `/app/models/my-yolo26m-model.onnx` (optional) — used for trtexec compilation - Runs both export methods (if files exist) and validates inference on each compiled engine using mydataset dataset. """ import logging import sys import subprocess import os from pathlib import Path from ultralytics import YOLO # Configure logging logging.basicConfig( level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s' ) logger = logging.getLogger(__name__) def run_yolo_checks(): """Execute yolo checks via CLI to log environment details.""" logger.info("=" * 60) logger.info("Environment Verification (yolo checks)") logger.info("=" * 60) try: res = subprocess.run(["yolo", "checks"], capture_output=True, text=True, check=True) logger.info("\n" + res.stdout.strip()) except subprocess.CalledProcessError as e: logger.error(f"yolo checks failed:\n{e.stderr}") except FileNotFoundError: logger.error("yolo CLI command not found. Ensure ultralytics is installed in this environment.") def test_export_to_engine(): models_dir = Path("/app/models") pt_path = models_dir / "my-yolo26m-model.pt" onnx_path = models_dir / "my-yolo26m-model.onnx" dataset_yaml = "/app/data/mydataset.yaml" logger.info("=" * 60) logger.info("Minimal OBB Test - Export & Validation") logger.info("=" * 60) engines_tested = [] # Part 1: Ultralytics native export (PT -> Engine) if pt_path.exists(): try: logger.info(f"\nRunning Ultralytics native export for: {pt_path.name}") model = YOLO(str(pt_path)) exported = model.export(format="engine", imgsz=1024, device=0, half=True) # Resolve the exported engine path native_engine = str(exported) if isinstance(exported, (str, Path)) else str(pt_path.with_suffix('.engine')) if os.path.exists(native_engine): logger.info(f"Ultralytics export successfully produced: {native_engine}") engines_tested.append(native_engine) else: logger.error(f"Ultralytics export completed, but engine file was not found at {native_engine}") except Exception as e: logger.error(f"Ultralytics native export failed: {e}", exc_info=True) else: logger.info(f"PT file not found ({pt_path}); skipping Ultralytics native export.") # Part 2: trtexec compilation (ONNX -> Engine) if onnx_path.exists(): try: logger.info(f"\nRunning trtexec to compile ONNX: {onnx_path.name}") trt_engine = str(onnx_path.with_suffix('.trt.engine')) cmd = [ "/usr/src/tensorrt/bin/trtexec", f"--onnx={onnx_path}", f"--saveEngine={trt_engine}", "--fp16", "--memPoolSize=workspace:768", # Prevents Out of Memory ] logger.info(f"Running command: {' '.join(cmd)}") res = subprocess.run(cmd, capture_output=True, text=True) if res.returncode != 0: logger.error("trtexec compilation failed.") logger.error(f"trtexec stderr:\n{res.stderr}") else: if os.path.exists(trt_engine): logger.info(f"trtexec successfully saved engine to {trt_engine}") engines_tested.append(trt_engine) else: logger.error(f"trtexec returned success, but engine file was not found at {trt_engine}") except Exception as e: logger.error(f"trtexec execution failed: {e}", exc_info=True) else: logger.info(f"ONNX file not found ({onnx_path}); skipping trtexec compilation.") if not engines_tested: logger.error("\nNo engines are available to test. Exiting validation phase.") return False # Part 3: Inference Validation validation_success = True for eng in engines_tested: try: logger.info(f"\nValidating inference with engine: {eng}") model = YOLO(eng, task="obb") metrics = model.val(data=dataset_yaml, imgsz=1024, batch=1, verbose=False) logger.info(f"Validation metrics for {eng}: mAP50-95={metrics.box.map:.4f}, mAP50={metrics.box.map50:.4f}") except Exception as e: logger.error(f"Validation failed on {eng}: {e}", exc_info=True) validation_success = False logger.info("\nExecution complete.") return validation_success if __name__ == "__main__": run_yolo_checks() try: success = test_export_to_engine() logger.info("\n" + "=" * 60) if success: logger.info("✓ Test passed") else: logger.info("✗ Test failed") logger.info("=" * 60) sys.exit(0 if success else 1) except Exception as e: logger.error(f"Unhandled script exception: {e}", exc_info=True) sys.exit(1)Output
obb_test_jp5 | Creating new Ultralytics Settings v0.0.6 file ✅ obb_test_jp5 | View Ultralytics Settings with 'yolo settings' or at '/root/.config/Ultralytics/settings.json' obb_test_jp5 | Update Settings with 'yolo settings key=value', i.e. 'yolo settings runs_dir=path/to/dir'. For help see https://docs.ultralytics.com/quickstart/#ultralytics-settings. obb_test_jp5 | 2026-05-14 12:51:35,392 - INFO - ============================================================ obb_test_jp5 | 2026-05-14 12:51:35,392 - INFO - Environment Verification (yolo checks) obb_test_jp5 | 2026-05-14 12:51:35,392 - INFO - ============================================================ obb_test_jp5 | 2026-05-14 12:51:50,554 - INFO - obb_test_jp5 | Ultralytics 8.4.50 🚀 Python-3.8.10 torch-2.1.0a0+41361538.nv23.06 CUDA:0 (Xavier, 6833MiB) obb_test_jp5 | Setup complete ✅ (2 CPUs, 6.7 GB RAM, 91.7/107.6 GB disk) obb_test_jp5 | obb_test_jp5 | OS Linux-5.10.216-tegra-aarch64-with-glibc2.29 obb_test_jp5 | Environment Docker obb_test_jp5 | Python 3.8.10 obb_test_jp5 | Install git obb_test_jp5 | Path /ultralytics/ultralytics obb_test_jp5 | RAM 6.67 GB obb_test_jp5 | Disk 91.7/107.6 GB obb_test_jp5 | CPU ARMv8 Processor rev 0 (v8l) obb_test_jp5 | CPU count 2 obb_test_jp5 | GPU Xavier, 6833MiB obb_test_jp5 | GPU count 1 obb_test_jp5 | CUDA 11.4 obb_test_jp5 | obb_test_jp5 | numpy ✅ 1.23.5>=1.23.0 obb_test_jp5 | matplotlib ✅ 3.7.5>=3.3.0 obb_test_jp5 | opencv-python-headless ✅ 4.13.0.92>=4.6.0 obb_test_jp5 | pillow ✅ 10.4.0>=7.1.2 obb_test_jp5 | pyyaml ✅ 6.0.3>=5.3.1 obb_test_jp5 | requests ✅ 2.32.4>=2.23.0 obb_test_jp5 | scipy ✅ 1.10.1>=1.4.1 obb_test_jp5 | torch ✅ 2.1.0a0+41361538.nv23.6>=1.8.0 obb_test_jp5 | torchvision ✅ 0.16.2+c6f3977>=0.9.0 obb_test_jp5 | psutil ✅ 7.2.2>=5.8.0 obb_test_jp5 | polars ✅ 1.8.2>=0.20.0 obb_test_jp5 | ultralytics-thop ✅ 2.0.19>=2.0.18 obb_test_jp5 | torch ✅ 2.1.0a0+41361538.nv23.6!=2.4.0,>=1.8.0; sys_platform == "win32" obb_test_jp5 | 2026-05-14 12:51:50,562 - INFO - ============================================================ obb_test_jp5 | 2026-05-14 12:51:50,563 - INFO - Minimal OBB Test - Export & Validation obb_test_jp5 | 2026-05-14 12:51:50,564 - INFO - ============================================================ obb_test_jp5 | 2026-05-14 12:51:50,565 - INFO - obb_test_jp5 | Running Ultralytics native export for: my-yolo26m-model.pt obb_test_jp5 | 2026-05-14 12:51:51,353 - ERROR - Ultralytics native export failed: ERROR ❌️ /app/models/my-yolo26m-model.pt requires numpy>=1.26.1, however numpy==1.23.5 is installed. obb_test_jp5 | Traceback (most recent call last): obb_test_jp5 | File "/ultralytics/ultralytics/nn/tasks.py", line 1466, in torch_safe_load obb_test_jp5 | ckpt = _load() obb_test_jp5 | File "/ultralytics/ultralytics/nn/tasks.py", line 1463, in _load obb_test_jp5 | return torch_load(file, map_location="cpu") obb_test_jp5 | File "/ultralytics/ultralytics/utils/patches.py", line 181, in torch_load obb_test_jp5 | return torch.load(*args, **kwargs) obb_test_jp5 | File "/usr/local/lib/python3.8/dist-packages/torch/serialization.py", line 877, in load obb_test_jp5 | return _load(opened_zipfile, map_location, pickle_module, **pickle_load_args) obb_test_jp5 | File "/usr/local/lib/python3.8/dist-packages/torch/serialization.py", line 1253, in _load obb_test_jp5 | result = unpickler.load() obb_test_jp5 | File "/usr/local/lib/python3.8/dist-packages/torch/serialization.py", line 1246, in find_class obb_test_jp5 | return super().find_class(mod_name, name) obb_test_jp5 | ModuleNotFoundError: No module named 'numpy._core' obb_test_jp5 | obb_test_jp5 | The above exception was the direct cause of the following exception: obb_test_jp5 | obb_test_jp5 | Traceback (most recent call last): obb_test_jp5 | File "src/minimal.py", line 58, in test_export_to_engine obb_test_jp5 | model = YOLO(str(pt_path)) obb_test_jp5 | File "/ultralytics/ultralytics/models/yolo/model.py", line 76, in __init__ obb_test_jp5 | super().__init__(model=model, task=task, verbose=verbose) obb_test_jp5 | File "/ultralytics/ultralytics/engine/model.py", line 144, in __init__ obb_test_jp5 | self._load(model, task=task) obb_test_jp5 | File "/ultralytics/ultralytics/engine/model.py", line 283, in _load obb_test_jp5 | self.model, self.ckpt = load_checkpoint(weights) obb_test_jp5 | File "/ultralytics/ultralytics/nn/tasks.py", line 1537, in load_checkpoint obb_test_jp5 | ckpt, weight = torch_safe_load(weight) # load ckpt obb_test_jp5 | File "/ultralytics/ultralytics/nn/tasks.py", line 1489, in torch_safe_load obb_test_jp5 | raise ModuleNotFoundError( obb_test_jp5 | ModuleNotFoundError: ERROR ❌️ /app/models/my-yolo26m-model.pt requires numpy>=1.26.1, however numpy==1.23.5 is installed. obb_test_jp5 | 2026-05-14 12:51:51,382 - INFO - obb_test_jp5 | Running trtexec to compile ONNX: my-yolo26m-model.onnx obb_test_jp5 | 2026-05-14 12:51:51,382 - INFO - Running command: /usr/src/tensorrt/bin/trtexec --onnx=/app/models/my-yolo26m-model.onnx --saveEngine=/app/models/my-yolo26m-model.trt.engine --fp16 --memPoolSize=workspace:768 obb_test_jp5 | 2026-05-14 13:21:39,214 - INFO - trtexec successfully saved engine to /app/models/my-yolo26m-model.trt.engine obb_test_jp5 | 2026-05-14 13:21:39,217 - INFO - obb_test_jp5 | Validating inference with engine: /app/models/my-yolo26m-model.trt.engine obb_test_jp5 | Ultralytics 8.4.50 🚀 Python-3.8.10 torch-2.1.0a0+41361538.nv23.06 CUDA:0 (Xavier, 6833MiB) obb_test_jp5 | Loading /app/models/my-yolo26m-model.trt.engine for TensorRT inference... obb_test_jp5 | [05/14/2026-13:21:41] [TRT] [I] Loaded engine size: 43 MiB obb_test_jp5 | [05/14/2026-13:21:45] [TRT] [I] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +260, GPU +306, now: CPU 648, GPU 5277 (MiB) obb_test_jp5 | [05/14/2026-13:21:46] [TRT] [I] [MemUsageChange] Init cuDNN: CPU +82, GPU +156, now: CPU 730, GPU 5433 (MiB) obb_test_jp5 | [05/14/2026-13:21:46] [TRT] [I] [MemUsageChange] TensorRT-managed allocation in engine deserialization: CPU +0, GPU +41, now: CPU 0, GPU 41 (MiB) obb_test_jp5 | [05/14/2026-13:21:46] [TRT] [I] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +0, GPU +28, now: CPU 687, GPU 5422 (MiB) obb_test_jp5 | [05/14/2026-13:21:46] [TRT] [I] [MemUsageChange] Init cuDNN: CPU +0, GPU +32, now: CPU 687, GPU 5454 (MiB) obb_test_jp5 | [05/14/2026-13:21:46] [TRT] [I] [MemUsageChange] TensorRT-managed allocation in IExecutionContext creation: CPU +0, GPU +92, now: CPU 0, GPU 133 (MiB) obb_test_jp5 | Setting batch=1 input of shape (1, 3, 1024, 1024) obb_test_jp5 | 2026-05-14 13:21:54,562 - INFO - generated new fontManager obb_test_jp5 | val: Fast image access ✅ (ping: 1.1±0.9 ms, read: 479.2±146.1 MB/s, size: 2563.4 KB) val: Scanning /app/data/labels/rec26.cache... 2940 images, 26 backgrounds, 0 corrupt: 100% ━━━━━━━━━━━━ 2940/2940 14.4Mit/s 0.0s obb_test_jp5 | 2026-05-14 13:22:01,033 - ERROR - Validation failed on /app/models/my-yolo26m-model.trt.engine: CUDA error: no kernel image is available for execution on the device obb_test_jp5 | CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. obb_test_jp5 | For debugging consider passing CUDA_LAUNCH_BLOCKING=1. obb_test_jp5 | Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions. obb_test_jp5 | Traceback (most recent call last): obb_test_jp5 | File "src/minimal.py", line 114, in test_export_to_engine obb_test_jp5 | metrics = model.val(data=dataset_yaml, imgsz=1024, batch=1, verbose=False) obb_test_jp5 | File "/ultralytics/ultralytics/engine/model.py", line 615, in val obb_test_jp5 | validator(model=self.model) obb_test_jp5 | File "/usr/local/lib/python3.8/dist-packages/torch/utils/_contextlib.py", line 115, in decorate_context obb_test_jp5 | return func(*args, **kwargs) obb_test_jp5 | File "/ultralytics/ultralytics/engine/validator.py", line 210, in __call__ obb_test_jp5 | model.warmup(imgsz=(1 if pt else self.args.batch, self.data["channels"], imgsz, imgsz)) # warmup obb_test_jp5 | File "/ultralytics/ultralytics/nn/autobackend.py", line 313, in warmup obb_test_jp5 | non_max_suppression(warmup_boxes) # warmup NMS obb_test_jp5 | File "/ultralytics/ultralytics/utils/nms.py", line 154, in non_max_suppression obb_test_jp5 | i = torchvision.ops.nms(boxes, scores, iou_thres) obb_test_jp5 | File "/usr/local/lib/python3.8/dist-packages/torchvision/ops/boxes.py", line 41, in nms obb_test_jp5 | return torch.ops.torchvision.nms(boxes, scores, iou_threshold) obb_test_jp5 | File "/usr/local/lib/python3.8/dist-packages/torch/_ops.py", line 646, in __call__ obb_test_jp5 | return self._op(*args, **kwargs or {}) obb_test_jp5 | RuntimeError: CUDA error: no kernel image is available for execution on the device obb_test_jp5 | CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. obb_test_jp5 | For debugging consider passing CUDA_LAUNCH_BLOCKING=1. obb_test_jp5 | Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions. obb_test_jp5 | obb_test_jp5 | 2026-05-14 13:22:01,521 - INFO - obb_test_jp5 | Execution complete. obb_test_jp5 | 2026-05-14 13:22:01,524 - INFO - obb_test_jp5 | ============================================================ obb_test_jp5 | 2026-05-14 13:22:01,526 - INFO - ✗ Test failed obb_test_jp5 | 2026-05-14 13:22:01,529 - INFO - ============================================================ obb_test_jp5 exited with code 1Export Device
This is a bit of info on where my ONNX files come from and what my PC environment looks like where I compiled them.
Yolo Checks on the ONNX Export Device
Ultralytics 8.4.48 🚀 Python-3.13.9 torch-2.11.0+cu130 CPU (AMD Ryzen 5 5600G with Radeon Graphics) Setup complete ✅ (12 CPUs, 15.0 GB RAM, 744.1/878.1 GB disk) OS Linux-6.17.0-23-generic-x86_64-with-glibc2.39 Environment Linux Python 3.13.9 Install git Path /home/myuser/projects/myproject/.venv/lib/python3.13/site-packages/ultralytics RAM 15.01 GB Disk 744.1/878.1 GB CPU AMD Ryzen 5 5600G with Radeon Graphics CPU count 12 GPU None GPU count None CUDA None numpy ✅ 2.4.4>=1.23.0 matplotlib ✅ 3.10.9>=3.3.0 opencv-python ✅ 4.13.0.92>=4.6.0 pillow ✅ 12.2.0>=7.1.2 pyyaml ✅ 6.0.3>=5.3.1 requests ✅ 2.33.1>=2.23.0 scipy ✅ 1.17.1>=1.4.1 torch ✅ 2.11.0>=1.8.0 torch ✅ 2.11.0!=2.4.0,>=1.8.0; sys_platform == "win32" torchvision ✅ 0.26.0>=0.9.0 psutil ✅ 7.2.2>=5.8.0 polars ✅ 1.40.1>=0.20.0 ultralytics-thop ✅ 2.0.19>=2.0.18Export Code
from ultralytics import YOLO # 1. Load the model weights model_path = "my-yolo26m-model.pt" model = YOLO(model_path) # 2. Export the model # Configured for OBB (Oriented Bounding Boxes) and legacy TensorRT compatibility success = model.export( format="onnx", imgsz=1024, # Target resolution opset=13, # ONNX operator set version, also tried 12 simplify=True, # I hoped this would help task="obb", half=False, # I tried this with True and False end2end=False # For legacy TensorRT support...? ) print(f"Export successful: {success}")YOLO26m-obb summary (fused): 175 layers, 23,443,236 parameters, 0 gradients, 71.5 GFLOPs PyTorch: starting from 'my-yolo26m-model.pt' with input shape (1, 3, 1024, 1024) BCHW and output shape(s) (1, 6, 21504) (90.8 MB) ONNX: starting export with onnx 1.21.0 opset 13... ONNX: slimming with onnxslim 0.1.93... ONNX: export success ✅ 2.5s, saved as 'my-yolo26m-model.onnx' (81.3 MB) Export complete (4.3s)Old Container
Jut in case it is relevant, this is the yolo checks of my self-made container which is based on l4t-ml.
root@xaviernx-desktop:/app# yolo checks WARNING ⚠️ user config directory '/root/.config/Ultralytics' is not writable, using '/tmp/Ultralytics'. Set YOLO_CONFIG_DIR to override. Creating new Ultralytics Settings v0.0.6 file ✅ View Ultralytics Settings with 'yolo settings' or at '/tmp/Ultralytics/settings.json' Update Settings with 'yolo settings key=value', i.e. 'yolo settings runs_dir=path/to/dir'. For help see https://docs.ultralytics.com/quickstart/#ultralytics-settings. Ultralytics 8.4.49 🚀 Python-3.8.10 torch-2.0.0+nv23.05 CUDA:0 (Xavier, 6833MiB) Setup complete ✅ (2 CPUs, 6.7 GB RAM, 84.9/107.6 GB disk) OS Linux-5.10.216-tegra-aarch64-with-glibc2.29 Environment Docker Python 3.8.10 Install pip Path /usr/local/lib/python3.8/dist-packages/ultralytics RAM 6.67 GB Disk 84.9/107.6 GB CPU ARMv8 Processor rev 0 (v8l) CPU count 2 GPU Xavier, 6833MiB GPU count 1 CUDA 11.4 numpy ✅ 1.23.5>=1.23.0 matplotlib ✅ 3.7.5>=3.3.0 opencv-python ❌ (not installed)>=4.6.0 pillow ✅ 10.2.0>=7.1.2 pyyaml ✅ 5.3.1>=5.3.1 requests ✅ 2.32.4>=2.23.0 scipy ✅ 1.10.1>=1.4.1 torch ✅ 2.0.0+nv23.5>=1.8.0 torch ✅ 2.0.0+nv23.5!=2.4.0,>=1.8.0; sys_platform == "win32" torchvision ✅ 0.15.1a0+42759b1>=0.9.0 psutil ✅ 5.9.6>=5.8.0 polars ❌ (not installed)>=0.20.0 ultralytics-thop ❌ (not installed)>=2.0.18
Murat RaimbekovThanks for the detailed MRE. I reproduced the numpy failure end-to-end with
torch.save/torch.loadand have a verified fix. There are two independent issues here.1.
numpy._core/ "requires numpy>=1.26.1" — root cause and fixYour
.ptwas saved on your PC with numpy 2.x / Python 3.13, so the checkpoint pickle embeds the numpy-2numpy._coremodule path. JetPack 5 on Xavier NX runs Python 3.8, whose maximum numpy is 1.24.x, and Xavier cannot move to JetPack 6 (Orin-only). I verified that even numpy 1.26.x does not fully read such a checkpoint, so "upgrade numpy" is not a real option on JP5.The reliable fix is a small compatibility shim that aliases
numpy._coreto the legacynumpy.core, injected before the model is loaded. I confirmed this loads a numpy-2torch.savecheckpoint correctly under numpy 1.23.5 with the model and all fields intact. Put this at the very top of yourmain.py, beforefrom ultralytics import YOLO:import sys, numpy, numpy.core sys.modules["numpy._core"] = numpy.core for _s in ("multiarray", "numeric", "umath", "_multiarray_umath"): try: sys.modules[f"numpy._core.{_s}"] = __import__(f"numpy.core.{_s}", fromlist=[_s]) except Exception: passAlternatively, fix the checkpoint once on the PC (where numpy 2.x can read it) so the device never needs the shim — load it and re-save it after converting the auxiliary numpy fields, then copy the new
.ptto the Jetson.2. Empty boxes /
CUDA error: no kernel image is available for execution on the deviceThese are separate from numpy:
- The
trtexec-built engine produces no detections because a rawtrtexecengine carries none of the Ultralytics export metadata, so the OBB decoding/post-processing is never applied. Build the engine on the Jetson with Ultralytics directly from the.pt(after applying the shim above) so metadata is embedded:YOLO("model.pt").export(format="engine", imgsz=1024, half=True, device=0). no kernel image is availableis raised insidetorchvision.ops.nmson the GPU: the PyTorch/torchvision inultralytics/ultralytics:latest-jetson-jetpack5is built for Orin (SM 8.7), while Xavier NX is SM 7.2. That is exactly why YOLO11 OBB works in your older l4t-ml-based container (torch 2.0.0+nv23.5) but the official image fails. Recommended path: stay on the container that already works on your Xavier,pip install ultralytics opencv-python-headlessthere, and run the Ultralyticsexport(format="engine")+val()entirely on-device.
TL;DR: the numpy error is a pickle module-path mismatch from the numpy-2 export PC (fix with the shim above — verified), and the empty/CUDA errors are a raw-
trtexecmetadata loss plus a torch/torchvision build that doesn't target Xavier's SM 7.2.1- The
RotatingGiraffe
Hi there, thank you for the help! I just recently got around to trying out the shim and now everything works just fine, thank you for your help!
yashpatel2010
How to run deepstream for YOLO26 models in my jetson orin nano developer kit?
Is there any documentation for deepstream on jetson? I don't want any installation errors of version missmatch.
So give me any perfect guide link so that I will not miss anything while installation and run inference on live RTSP streams
githubwalt(editado)
Here you go: https://docs.ultralytics.com/guides/deepstream-nvidia-jetson/
Have a look in jtop (info) for your board compute capability (CUDA Arch BIN). If it's also 8.7, then see my posts here above to avoid problems with torch.
Glenn JocherUltralyticsAdding one detail for
Jetson Orin Nano: keep the versions paired asJetPack 6.1+DeepStream 7.1from the Jetson quick-start and NVIDIA DeepStream getting started; for live RTSP, set theuriindeepstream_app_config.txtto your stream, and if you hit any mismatch please post the full text traceback plusyolo checksoutput here so the team can help.yashpatel2010
Thanks for your reply. Just onemore doubt to clear. docker pull nvcr.io/nvidia/deepstream:7.1-triton-multiarch
if I use above docker container for my jetpack6.1 so it there is no issue of version missmatch of deepstream right?
is it works very well with Yolo26 deepstream settings which is showing in this doc? https://docs.ultralytics.com/guides/deepstream-nvidia-jetson/
what is your recommendation to run Multiple RTSP streams with deepstream with docker containers or without docker containers manual installation?
Glenn JocherUltralyticsYes,
nvcr.io/nvidia/deepstream:7.1-triton-multiarchis the correct DeepStream 7.1 Jetson container forJetPack 6.1, but it still depends on the host having the JetPack BSP and NVIDIA container runtime installed, so it lowers mismatch risk rather than removing it completely; for YOLO26 + RTSP, the DeepStream multi-stream section is the right reference, and I’d suggest native setup first since that’s the install path used in the guide, then move to Docker once your pipeline is stable. (docs.ultralytics.com)yashpatel2010(editado)
Thanks. For native setup Is it good to install deepstream 7.1 while installing jetpack6.1 only?
Glenn JocherUltralyticsYes—on
Jetson Orin Nanonative setup,JetPack 6.1+DeepStream 7.1is the correct pairing; installJetPack 6.1first, boot and verify the system, then installDeepStream 7.1separately using the versions listed in the Jetson + YOLO26 guide and the DeepStream on Jetson guide to avoid mismatches.
githubwalt
Installing torch 2.11.0 from the official website pytorch.org on an AGX Orin 64Gb (dev kit) with JP6.2.2 or installing torch 2.10.0 and torchvision 0.25.0 following this page, give complaints about not being compatible with the Compute Capability (CC) of my machine which has 8.7, whatever that means. Only installing the previous versions torch 2.8.0 and torchvision 0.23.0 do work.
I understand that torch needs to be tailored specifically for the Jetson family and that is what this page is all about, so why do I have this issue if I follow the instructions given here? Is it because I am on Jetpack 6.2.2 and here it's 6.1?
githubwalt
I forgot the warning and error output when running a python test script:
UserWarning: Found GPU0 Orin which is of compute capability (CC) 8.7.The following list shows the CCs this version of PyTorch was built for and the hardware CCs it supports:- 8.0 which supports hardware CC >=8.0,<9.0 except {8.7}- 9.0 which supports hardware CC >=9.0,<10.0
torch.AcceleratorError CUDA error: no kernel image is available for execution on the device
Glenn JocherUltralyticsThanks, that traceback helps —
CC 8.7is the AGX Orin GPU architecture, andno kernel image is availablemeans the installed PyTorch wheel was built withoutsm_87support. So yes, this is most likely a JetPack/PyTorch wheel mismatch: the native steps in the JetPack 6.1 section don’t necessarily map cleanly toJP6.2.2, even though the page includesJP6.2benchmark coverage. For now I’d stick with the workingtorch 2.8.0+torchvision 0.23.0pair or use the JetPack 6 Docker setup; if you want us to check your exact environment, please postyolo checksand the full text traceback here in the discussion.githubwalt
Thanks, will stick to 2.8 🙂 👍 (I also posted on the guide page https://docs.ultralytics.com/guides/nvidia-jetson/#what-are-the-best-practices-for-maximizing-performance-on-nvidia-jetson-when-using-yolo26 , so you could ignore that if you like, can't delete my post there 😏 )
Glenn JocherUltralyticsNo worries — the duplicate is fine, and if you later confirm a stable
JP6.2.2+ PyTorch pairing on AGX Orin, please add it here so we can reflect it in the Jetson guide.githubwalt
Will do. Now trying to get the most out of the AGX. Experimenting with number of workers on a batch size of 16. So far 12 workers gave memory issues ... I chose 12 because of 12 cores, but that was a little too much. Now running with 8 workers ... so far so good. To be continued 🙂
githubwalt
Hi Glenn, I can confirm a stable JP6.2.2 + Pytorch (torch 2.8.0 and torchvision 0.23.0) on the AGX Orin.